Efficient combination of deep learning and tree-based classification models for solar panel dust detection

Solar panels are crucial for converting sunlight into electricity. However, their efficiency and performance can significantly decline due to environmental factors, notably the buildup of dust and debris on their surfaces. This study proposes a hybrid model comprising a deep learning component for feature extraction and tree-based classifiers specifically designed to distinguish between ”dusty” and ”clean” solar panels. The objective is to develop a robust model to accurately detect dust on solar panels under various environmental conditions. Our approach leverages pre-trained deep learning models fine-tuned to detect dust on photovoltaic panels to extract relevant features. These features are then used for classification using lightweight tree-based models. Fine-tuning the pre-trained model weights significantly improves the detection performance. The results show that the combination of features extracted from EfficientNetB7 and a vision transformer achieves the highest accuracy of 97% when fed into a tree classifier. In addition, introducing a tree-based model improves all classification metrics compared to a fully dense connected layer. This work can be adapted to detect dust levels and, consequently, to help identify effective cleaning methods in an automated manner. • Enhancing dataset quality and model performance through data cleaning. • Comparing fine-tuned deep neural networks for feature extraction. • Identifying the best tree-based model for dust image classification. • Building lightweight classifiers through feature selection. • Developing a solar panel dust detection system using a hybrid machine learning solution.